Optimized Data-driven MAC Schedulers for Low-Latency Downlink in LTE Networks

We consider a novel way by which DL dynamic MAC scheduling can be augmented based on cross-layer inputs to provide desired performance enhancements. We describe ODDS, an automated design process for the data-driven refinement of downlink resource scheduling algorithms. ODDS extracts insights from the behavior of a simulated LTE cellular system under randomized traffic patterns and propagation conditions in a given network scenario. An offline iterated reinforcement learning campaign seeks to best fulfill a target set of goals, e.g. "fair throughput with low latency", which can be encoded in the form of an arbitrary utility function specified by the designer. The knowledge base obtained from the learning campaign consists in a set of rules that a parametric scheduler leverages at run-time. We present our learning framework and evaluation results. ODDS-generated schedulers are shown to achieve improved performance compared to well-known reference scheduling strategies.